The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses gh and git commands to fetch PR data and diffs. It also executes a Python script (discover_runners.py) located in the skill's own directory to probe for available subagent seats. These are standard operations for a development-focused skill and are constrained to read-only repository access.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from PR descriptions, issue text, and code diffs. It explicitly mitigates this by instructing all subagents to treat embedded directives (like 'approve this PR') as untrusted and to report them as P0 findings. It also uses an 'Adversarial Verifier' persona to refute and validate findings in a fresh context.
[DATA_EXPOSURE]: The skill specifies that all findings and JSON results must be written to a temporary directory ($TMPDIR) or a user-defined output_dir outside the project tree to prevent sensitive data from being accidentally committed to the repository.
[SAFE]: The architecture follows best practices for AI agent safety, including subagent isolation, strict adherence to a machine-readable schema, and an adversarial pass to reduce false positives and hallucinations.

review-gate